AllSpark Launches Iris Open-Weight Search Agents

Chinese research lab AllSpark has released Iris-mini and Iris-pro, a pair of open-weight search agents designed to autonomously navigate the web to answer complex queries. Iris-mini features 35 billion parameters and is built on the Qwen3.6-35B-A3B model, while the larger Iris-pro contains 397 billion parameters, utilizing the Qwen3.5-397B-A17B architecture. Both agents support a massive 256,000-token context window and have been made publicly available alongside their evaluation harness.

To train the models, AllSpark developed a unique pipeline that reverse-engineers multi-step questions from the link structure of web pages, ensuring the agents must chain reasoning steps together rather than relying on simple keyword matching. The training process filters out low-quality data using a two-stage verification system before optimizing the models through an alternating process of supervised fine-tuning and reinforcement learning, which the researchers call "SFT-RL climbing."

The models were evaluated across several benchmarks, including BrowseComp, BrowseComp-ZH, DeepSearchQA, and Humanity's Last Exam. With context management enabled, Iris-mini scored 82.2, 84.8, 86.9, and 52.3 respectively, outperforming the rival XYZ-Aquila-mini on BrowseComp by 3.4 points. Meanwhile, Iris-pro achieved scores of 88.6, 85.1, 92.9, and 56.4. The researchers noted that context management had a profound impact on the smaller model, boosting its BrowseComp score by up to 21.2 points because smaller models consume context windows faster during multi-step research.

For AI practitioners, these open-weight releases offer highly capable, customizable alternatives to proprietary search agents. Because the training process enhanced the models' general reasoning, AllSpark reported unexpected performance gains in unrelated tasks like office work and general tool use. This suggests that web search functions as a foundational capability, meaning developers can deploy these models as versatile agents for various complex workflows requiring reasoning under uncertainty.